Run AI agents on an ad batch: briefs and boundaries

Divide the batch by outputs

A useful agent assignment ends in something another person can inspect: an approved script draft, a voice asset, a shot list or a validated render. “Help with the campaign” does not define a completion boundary. Split the batch at those boundaries so the next worker receives a defined result rather than an unfinished conversation.

Anthropic’s subagent documentation describes separate working contexts. Its skills guide describes reusable instructions. Use the first for bounded workers and the second for the method they follow. claude for ads provides the wider context, while this guide focuses on keeping a multi-asset batch understandable when several jobs run at once.

In a hypothetical desk lamp batch, the writing agent can finish its assignment while a product fact remains unresolved. Its handoff should identify that question instead of silently choosing an answer. The reviewer checks the draft against the brief; the render worker waits for accepted inputs. These boundaries make responsibility inspectable without requiring every agent to read the entire production conversation or infer which draft became authoritative.

Each assignment needs input paths, the permitted operation, output names, content rules and a definition of done. A voice assignment might require a standard script, pronunciation script, playable audio and a pronunciation review note. It should not authorize changing the approved product statement.

Decide who resolves ambiguous inputs. A worker should stop that specific asset when two references disagree, while continuing independent assets that are clear. This is different from halting a whole batch whenever one detail is unresolved. Write the stop rule into the brief so the worker does not choose between guessing and abandoning everything. The merge owner then knows which decision remains with them.

Source: Turntake editorial checklist (proposed), October 2026.

Give every worker an isolated output area

In our production in October 2026, parallel workers write in separate folders and do not modify shared rendering scripts. Isolation protects both the work and the evidence: the file in a worker’s directory can be traced to that assignment without wondering whether another worker overwrote it.

1. Clay
2. 3D
3. Crochet

Keep the source library readable and provide a copied script when experimentation is required. Name outputs with the asset ID and revision rather than a generic final name. A worker can produce several candidates, but it should mark which is proposed for review. The merge owner should never have to infer the intended deliverable from modification time or the order in a directory listing.

1. 01List distinct assetsAssign one owner each
2. 02Write acceptance briefFix inputs and limits
3. 03Produce in isolated folderPreserve shared sources
4. 04Inspect actual outputsMerge accepted revisions

Put limits beside the tool authority

Set a hard cap for any paid work, a limit on concurrent jobs and a disk allowance for temporary media. Record submitted request IDs as work happens. Limits should be operational inputs, not a sentence hidden at the end of a long conversation.

Our October 2026 production rules keep paid generation and account authority separate from an editing worker. Apply the same principle to each assignment. A tagging worker needs readable media and an output table; it does not need an ad-account mutation tool. A renderer needs the recipe and assets; it does not need to generate replacement footage when a source is missing. Escalate the missing asset to its owner.

Merge actual files through a checklist

Count deliverables from the filesystem and compare them with the assignment manifest. A worker message saying “all done” is useful context, but the merge gate checks the actual files. Verify names, expected formats, required fields and whether the outputs refer to existing assets.

For a render, inspect media streams and duration as well as file existence. For a script, inspect claims and language. For a tag table, inspect allowed vocabulary and uncertainty fields. A checker that does not fix: review AI ad outputs explains independent review. Keep defects linked to the asset and return them to the author, rather than letting the merger quietly rewrite files and blur who validated the final version.

Worker boundaryMerge evidence

• Worker boundaryScript draftMerge evidenceClaims and language checked
• Worker boundaryVoice assetMerge evidenceAudio and script comparison
• Worker boundaryRender outputMerge evidenceStreams, duration and visual review

Track dependencies without sharing mutable drafts

A motion worker can start after its still is approved. A caption worker can start after the final voice exists. Those are explicit dependencies. Letting both workers operate on a file that is still changing creates avoidable confusion about which input each result used.

Use immutable input IDs in the assignment record. When an upstream asset changes, identify the downstream jobs that need another pass. Do not rerun unrelated assets merely because the batch has a new revision. AI project memory: decisions in files, not chat covers keeping current task state and durable decisions in files. A video edit JSON recipe that survives revision covers the edit recipe that binds downstream renders to their exact ingredients.

End with an accountable batch manifest

The batch owner delivers a manifest of accepted outputs, rejected candidates and unresolved items. It should be possible to tell which files passed mechanical checks, which received human review and which are still proposals. Avoid a single status label that hides those differences.

Keep the report compact: assignment, asset ID, review status, remaining issue and next owner. If an agent stopped early, reassign the missing outputs from that manifest. Do not rely on a long chat transcript to reconstruct the gap.

Turntake’s model plans describes Turntake’s model plans. The parallel workflow here is a studio practice, not a claim that Turntake already launches an autonomous workforce. Its value comes from the written boundaries: each worker knows its inputs, each reviewer knows its evidence and one owner remains responsible for merging a coherent final batch.

Recover only the missing assignments

Suppose a worker delivers scripts A and B but not C. Mark the first two as awaiting their normal review and reassign C with the original brief. Do not reopen A and B simply because the worker stopped. If the brief itself was defective, identify which accepted outputs are affected before replacing them. This is a recommended recovery method. It keeps incomplete execution separate from invalid inputs, which are different reasons for returning work to an author.

Who wrote this We made the ads first, for SmartSolar in Denmark, Sweden and Ireland. Then we built Turntake. Where Turntake comes from